gradient boosted regression tree

Estimating crop evapotranspiration (ETc) in arid and semi-arid areas can be difficult due to the dynamic nature of this process across both time and space. In addition, obtaining on-site measurements for this variable can be very time-consuming and costly. This study aimed to develop a framework that accurately estimates the sugarcane crop evapotranspiration on a spatio-temporal scale. This was achieved using four machine learning (ML) algorithms (MLR, CART, SVR, and GBRT) combined with remote sensing (RS) data and meteorological variables. Also, to reduce the dependence on several meteorological parameters in conventional ETc equations, the performance of eight different experimental temperature-based methods and four modified Hargreaves & Samani equations was evaluated compared to the standard FAO-Penman-Monteith method. For this purpose, weather data were collected from Hakim Farabi Sugarcane Agro-Industrial meteorological station for three years (2018-2021). Nine combinations of input variables (RS data and meteorological variables) were designed based on the IGR method and then evaluated by the ML algorithms. The results showed that the highest accuracy of ML algorithms based on R2, RMSE, and MAE statistics was obtained in CART (0.99, 0.41, and 0.18) and GBRT algorithms (0.99, 0.65, and 0.26), respectively. Regarding temperature-based methods, Ivanov’s equation had the best performance with an R2 of 0.91, while Baier and Robertson’s equation had the weakest performance with an R2 of 0.78 when estimating ETc. Overall, the combination of RS and ML algorithms effectively produced more precise and reliable ETc values on both temporal and spatial scales.